Skip to main content

Über dieses Buch

This book constitutes the thoroughly refereed proceedings of the Second IEEE Colombian Conference, ColCACI 2019, held in Barranquilla, Colombia, in June 2019.
The 21 full papers presented were carefully reviewed and selected from 59 submissions. The papers cover such topics as video processing; biomedical systems; image processing, etc.



Video Processing


SVM and SSD for Classification of Mobility Actors on the Edge

In the context of video processing, transmission to a remote server is not always possible nor suitable. Video processing on the edge could offer a solution. However, lower processing capacities constraint the number of techniques available for devices, in this work we report the performance of two techniques for classification from video on a minicomputer. The implementation of a real-time vehicle counting and classification system is evaluated through Support Vector Machine (SVM) and the Single Shot Detector Framework (SSD) in a minicomputer. We compare two SVM bases techniques, IPHOG and a MBF using Scale Invariant Features. The obtained results show that with a video resolution of 1280 \(\times \) 720 pixels and using SVM, precision and recognition rates of 86% and 94% are obtained respectively, while with SSD 93% and 67% rates are reached with times of processing higher than SVM.
Andrés Heredia, Gabriel Barros-Gavilanes

Framework Comparison of Neural Networks for Automated Counting of Vehicles and Pedestrians

This paper presents a comparison of three neural network frameworks used to make volumetric counts in an automated and continuous way. In addition to cars, the application count pedestrians. Frameworks used are: SSD Mobilenet re-trained, SSD Mobilenet pre-trained, and GoogLeNet pre-trained. The evaluation data set has a total duration of 60 min and comes from three different cameras. Images from the real deployment videos are included when training to enrich the detectable cases. Traditional detection models applied to vehicle counting systems usually provide high values for cars seen from the front. However, when the observer or camera is on the side, some models have lower detection and classification values. A new data set with fewer classes reach similar performance values as trained methods with default data sets. Results show that for the class cars, recall and precision values are 0.97 and 0.90 respectively in the best case, making use of a trained model by default, while for the class people the use of a re-trained model provides better results with precision and recall values of 1 and 0.82.
Galo Lalangui, Jorge Cordero, Omar Ruiz-Vivanco, Luis Barba-Guamán, Jessica Guerrero, Fátima Farías, Wilmer Rivas, Nancy Loja, Andrés Heredia, Gabriel Barros-Gavilanes

Retail Traffic-Flow Analysis Using a Fast Multi-object Detection and Tracking System

Traffic-flow analysis allows to make critical decisions for retail operation management. Common approaches for traffic-flow analysis make use of hardware-based solutions, which have major drawbacks, such as high deployment and maintenance costs. In this work, we address this issue by proposing a Multiple-Object Tracking (MOT) system, following the tracking-by-detection paradigm, that leverages on an ensemble of detectors, each running every f frames. We further measured the performance of our model in the MOT16 Challenge and applied our algorithm to obtain heatmaps and paths for customers and shopping carts in a retail store from CCTV cameras.
Richard Cobos, Jefferson Hernandez, Andres G. Abad

Accelerating Video Processing Inside Embedded Devices to Count Mobility Actors

The actual number of surveillance cameras and the different methods for counting vehicles originate the question: What is the best place to process video flows? This work analyze techniques to accelerate a counting system for mobility actors like cars, pedestrians, motorcycles, bicycles, buses, and trucks in the context of an Edge computing application using deep learning. To solve this problem this study presents the analysis and implementation of different techniques based on the use of an additional hardware element as is the case of a Vision Processing Unit (VPU) in combination with methods that affect the resolution, bit rate, and time of video processing. For this purpose we consider the Mobilenet-SSD model with two approaches: a pre-trained model with known data sets and a trained model with images from our specific scenarios. Additionally, we compare an optimized model using OpenVINO toolkit and overclock of hardware. The use of SSD-Mobilenet’s model generates different results in terms of accuracy and time of video processing in the system. Results show that the use of an embedded device in combination with a VPU and video processing techniques reach 18.62 Frames per Second (FPS). Thus, video processing time is slightly superior (5.63 min) for a video of 5 min. Optimized model and overclock show improvements too. Recall and precision values of 91% and 97% are reported in the best case (class car) for the vehicle counting system.
Andrés Heredia, Gabriel Barros-Gavilanes

Biomedical Systems


Influences of the Trained State Model into the Decoding of Elbow Motion Using Kalman Filter

The properties of the Kalman Filter to decode elbow movement from non-invasive EEG are analyzed in this article. A set of configuration parameters using cross-validation are tested in order to find the ones that reduce the estimation error. Found that selecting correctly the number of channels and the time step used to configure the signal, it is possible to improve the filter estimation capabilities. As there was an apparent incidence of the variations in the recorded data used to train the model, an investigation of how those alterations affect the estimation precision in various data sets was made. The presented results showed that significant variations in the velocity and acceleration of the data set trains filters with lower accuracy than the ones built from a more uniform set.
E. Y. Veslin, M. S. Dutra, L. Bevilacqua, L. S. C. Raptopoulos, W. S. Andrade, J. G. M. Soares

Building Dengue Sensors for Brazil Using a Social Network and Text Mining

The increasing use of Social Networks to share personal information has opened many resources to analyze the behaviour of one city, state or country. Topics related to politics, science, health alarms and others are shared by users everyday to monitor an prevent events. Natural Language Processing is the tool to analyze this text and get some insight using Machine Learning Techniques. In this work, Twitter is analyzed to detect social events because users are considered sensors. The analysis is performed over Brazilian tweets to detect dengue. The results show the utility of the proposal to recognize dengue epidemics in the Brazilian territory.
Josimar Edinson Chire Saire

Prediction of Diabetic Patient Readmission Using Machine Learning

Hospital readmissions pose additional costs and discomfort for the patient and their occurrences are indicative of deficient health service quality, hence efforts are generally made by medical professionals in order to prevent them. These endeavors are especially critical in the case of chronic conditions, such as diabetes. Recent developments in machine learning have been successful at predicting readmissions from the medical history of the diabetic patient. However, these approaches rely on a large number of clinical variables thereby requiring deep learning techniques. This article presents the application of simpler machine learning models achieving superior prediction performance while making computations more tractable.
Juan Camilo Ramírez, David Herrera

A Computational Model for Distance Perception Based on Visual Attention Redeployment in Locomotor Infants

Self-locomotion experience of infants has been argued to improve perception of distance, as visual attention is drawn to previously undetected or ignored depth specifying information. We present a computational model to evaluate how does self-locomotion experience influences the estimation of distance in infants. The model assigns an estimated distance label to salient objects in the scene, through a Binocular Neural Network (BNN) that computes binocular disparities. Emphasizing on key aspects of locomotion experience, two BNN are trained, one for non-locomotor infants and one for locomotor infants. The validation and test stages of the process show a significant improvement on the distance estimation task for the BNN trained with locomotor experience. This result is added to previous evidence which supports that locomotion in infants is an important step in cognitive development.
Liz A. Jaramillo–Henao, Adrián A. Vélez–Aristizábal, Jaime E. Arango–Castro

DCGAN Model Used to Generate Body Gestures on a Human-Humanoid Interaction System

The current availability of the humanoid robots opens up a wide range of applications, for instance, in the domain of hospitality the humanoids can be programmed to behave autonomous ways to provide help to people. The aspect of the humanoids and the humanness of interaction are key components of success. We developed a system to endow the humanoid robot Pepper, from SoftBank Robotics, with the capability of both: identify the emotion state of the humans and exhibiting emotional states via gestures and postures generated using a DCGAN model to learn from the human body language and to create originals body expressions to exhibit like-human movements.
Francisco J. González, Andres Perez-Uribe, Hector F. Satizábal, Jesús A. López

Social Impact Assessment on a Mining Project in Peru Using the Grey Clustering Method and the Entropy-Weight Method

Social impact assessment (SIA) has become a key factor for environmental conflicts prevention. In this study, we conducted SIA using the grey clustering method, which is based on grey systems theory. In addition, the entropy-weight method was applied to determine the most important criterion that could generate a social conflict. A case study was conducted on a mining project located in Cajamarca, Peru. Two stakeholder groups and seven evaluation criteria were identified. The results revealed that for the urban population stakeholder group, the project would have positive social impact; and for the rural population stakeholder group, the project would have negative social impact. In addition, the criterion C7 (Access to drinking water rate per year in the department of Cajamarca) was the most import criterion that could generate a social conflict. Consequently, the method showed quantitative and qualitative results that could help central and local governments to make the best decision on the project. Moreover, the grey clustering method showed interesting results that could be apply to assess other projects or programs from social point of view.
Alexi Delgado, Chiara Carbajal, H. Reyes, I. Romero

Image Processing


Classification of Satellite Images Using Rp Fuzzy C Means

The computational capacities increase, the decrease of equipment costs, the growing need for information, among other reasons; It makes possible the increasingly common access to satellite data. In this context. The investigation of techniques related to remote sensing becomes very important because it provide important information about the earth’s surface. Currently, segmentation is an essential step in applications that make use of satellite images. However, the main problem is: “the data in a multispectral image shows a low statistical separation and a long quantity of data”. In this article we propose to improve the balancing of elements for the clusters. We use a new term to estimate the influence that each element must have for the each cluster. This new term can be understood as a repulsion factor and aims to increase the differences between groups. This modification is inspired by new term that was integrated into the NFCC algorithm (New Fuzzy Centroid Cluster).
For the tests, we use the internal validity of the cluster to compare the algorithms. Using the index we measure the characteristics of the segmentation and corroborate them with the final visual results. Therefore, we conclude that the addition of this new term allows balancing the elements for each group. As a result we conclude that the new term organizes the elements better because it avoids a fast convergence of the algorithm. Finally, the results show that this new factor generates clusters with lower entropy and greater similarity between the elements.
Luis Mantilla

A Feature Extraction Method Based on Convolutional Autoencoder for Plant Leaves Classification

In this research, we present an approach based on Convolutional Autoencoder (CAE) and Support Vector Machine (SVM) for leaves classification of different trees. While previous approaches relied on image processing and manual feature extraction, the proposed approach operates directly on the image pixels, without any preprocessing. Firstly, we use multiple layers of CAE to learn the features of leaf image dataset. Secondly, the extracted features were used to train a linear classifier based on SVM. Experimental results show that the classifiers using these features can improve their predictive value, reaching an accuracy rate of 94.74%.
Mery M. Paco Ramos, Vanessa M. Paco Ramos, Arnold Loaiza Fabian, Erbert F. Osco Mamani

Acquisition System Based in a Low-Cost Optical Architecture with Single Pixel Measurements and RGB Side Information for UAV Imagery

Spectral imaging has a wide range of applications, like remote sensing or biomedical imaging. In recent years, with the increasing use of unmanned aerial vehicles (UAVs), it is common to attach a spectral camera to an UAV to acquire information. Unfortunately, the spectral cameras used with UAV are expensive. Therefore this work proposes a low-cost optical architecture to acquire spectral images. The proposed architecture takes advantage of the UAV movement to spectral imaging. The proposed optical architecture is composed of two arms, one has an RGB camera, and the other a single-pixel spectrometer. The results show that depending on the sensing conditions selected, it is possible to retrieve high-quality spectral images. Simulation results show that the proposed architecture improves image quality in terms of PSNR compared with an RGB camera and the single-pixel camera (SPC), up to 1.31 dB and 20.44 dB respectively, also obtained a performance similar to an architecture that combine the SPC and RGB, and even besting it improving the quality of the image in terms of PSNR up to 0.49 dB. Moreover, the optical architecture proposed has the advantage of reducing the amount of sensed information in comparison to SPC and SPC + RGB; also, the implementation costs are reduced drastically because the proposed architecture does not use a digital micromirror device (DMD) to codify the incoming scene, which is the case of the SPC and SPC + RGB architectures.
Arnold Guarin, Homero Ortega, Hans Garcia

Integration of an Adaptive Cellular Automaton and a Cellular Neural Network for the Impulsive Noise Suppression and Edge Detection in Digital Images

This article proposes the combination of two bio-inspired computational models that are sequentially implemented to eliminate impulsive noise and edge detection in grayscale images. In general, this procedure consists of: (1) implementing a cellular automaton (CA) with an adaptive behavior that expands the Moore neighborhood when it considers that the information obtained from its first level neighbors is insufficient. Based on the above, the image affected by noise is processed, in order to eliminate the corrupted pixels and perform reprocessing that will lead to the improvement of the quality of the image, (2) the resulting image is defined as an input of the cellular neural network (CNN) together with the training images, so that by defining three templates (feedback (\( A \)), cloning (\( B \)) and threshold or bias (\( I \))), contour detection of objects within the image thrown by the initial method can be performed. The results for the noise elimination present a restoration of the image that oscillates between \( 70.63\% \) and \( 99.65\% \), indicating that the image does not lose its quality despite being exposed to high noise levels, similarly it occurs for the edge detection, which presents an approximate efficiency of \( 65\% \) with respect to the algorithms established within the framework of comparison.
Karen Angulo, Danilo Gil, Helbert Espitia

Sweet Citrus Fruit Detection in Thermal Images Using Fuzzy Image Processing

In agriculture, intelligent systems applications have generated great advances in automating some processes in the production chain. To improve the efficiency of those systems is proposes a vision algorithm to estimate the amount of fruits in sweet orange trees. This study proposes a computer vision system based on the capture of thermal images and fuzzy image processing. A bibliographical review has been done to analyze the state-of-the-art of the different systems used in fruit recognition, and also the different applications of thermography in agricultural systems. The algorithm developed for this project uses the intensification operator to contrast-enhanced and the fuzzy divergence for segmentation and Hough transform for fruit identification. It estimates the numbers of fruits in the tree, a task that is currently manually performed. In order to validate the proposed algorithm a database was created with images of sweet orange acquired in the Maringá Farm. The validation process indicated that the variation of the tree branch and the fruit temperature is not very high, making it difficult to segment the images using a temperature threshold. Errors in the segmentation algorithm could mean the increase of false positives in the fruit-counting algorithm. Recognition of isolated fruits with the proposed algorithm presented an overall accuracy of 93.5% and grouped fruits accuracy was 80%. The experiments show the need of other image hardware to improve the recognition of small temperature changes in the image.
Ingrid Lorena Argote Pedraza, John Faber Archila Diaz, Renan Moreira Pinto, Marcelo Becker, Mario Luiz Tronco



Bayesian Inference for Training of Long Short Term Memory Models in Chaotic Time Series Forecasting

For time series forecasting, obtaining models is based on the use of past observations from the same sequence. In those cases, when the model is learning from data, there is not an extra information that discuss about the quantity of noise inside the data available. In practice, it is necessary to deal with finite noisy datasets, which lead to uncertainty about the propriety of the model. For this problem, the employment of the Bayesian inference tools are preferable. A modified algorithm used for training a long-short term memory recurrent neural network for time series forecasting is presented. This approach was chosen to improve the forecasting of the original series, employing an implementation based on the minimization of the associated Kullback-Leibler Information Criterion. For comparison, a nonlinear autoregressive model implemented with a feedforward neural network was also presented. A simulation study was conducted to evaluate and illustrate results, comparing this approach with Bayesian neural-networks-based algorithms for artificial chaotic time-series and showing an improvement in terms of forecasting errors.
Cristian Rodríguez Rivero, Julián Pucheta, Daniel Patiño, Jose Luis Puglisi, Paula Otaño, Leonardo Franco, Gustavo Juarez, Efrén Gorrostieta, Alvaro David Orjuela-Cañón

Entropy and PCA Analysis for Environments Associated to Q-Learning for Path Finding

This work is based on the simulation of the reinforcement learning method for path search for mobile robotic agents in unstructured environments. The choice of the learning and reward coefficients of the Q-learning method affect the number of average actions that the algorithm requires for reach the goal from the start position. In addition, another important factor is the randomness degree and environment size over which the path must be calculated, since they affect the time of convergence in the learning. Likewise, a performance metric of the Q-learning algorithm is proposed, based on the Entropy and Principal Component Analysis of the environment representative images. The analysis by Entropy only allows to determine, in a scalar form, the environment randomness degree, but it does not provide information about the dispersion location. In contrast, the analysis by PCA allows to quantify not only the randomness, but also helps to estimate the direction of greater randomness of the environment. The advantage of this analysis by PCA and Entropy is that one could estimate the actions number or movements required for path search algorithms based on the randomness of unstructured environments.
Manuel Garcia-Quijada, Efren Gorrostieta-Hurtado, Jose Emilio Vargas-Soto, Manuel Toledano-Ayala

Fail Detection in WfM/BPM Systems from Event Log Sequences Using HMM-Type Models

Currently, there is an increasing interest in predicting the behavior of active work items in Business Process Management (BPM) systems, which would make possible to monitor the behavior of such processes in a more accurate way. Given the complexity of current business processes, conventional techniques are not always effective in addressing this type of requirements; therefore, machine learning techniques are being increasingly more used for this task. This work deals with the problem of fail detection in a BPM system from event logs, based on machine learning methods. The paper explores the use of three structural learning models, Hidden Markov Models (HMM), Hidden semi-Markov models (HSMM) and Non-stationary Hidden semi-Markov models (NHSMM). The experiments are carried out using a real database of about 460,000 event logs sequences. The results show that for the given dataset, fail detection can be achieved with an accuracy of 86.70% using the HSMM model. In order to reduce the computational load of the proposed approach, the models were implemented in a distributed processing environment using Apache Spark, which guarantees solution scalability.
Johnnatan Jaramillo, Julián D. Arias-Londoño

Solar Spots Classification Using Pre-processing and Deep Learning Image Techniques

Machine learning techniques and image processing have been successfully applied in many research fields. Astronomy and Astrophysics are some of these areas. In this work, we apply machine learning techniques in a new approach to classify and characterize solar spots which appear on the solar photosphere which express intense magnetic fields, and these magnetic fields present significant effects on Earth. In our experiments we consider images from Helioseismic and Magnetic Imager (HMI) in IntensitygramFlat format. We apply pre-processing techniques to recognize and count the groups of sunspots for further classification. Besides, we investigate the performance of the CNN AlexNet layer input in comparison with the Radial Basis Function Network (RBF) using different levels and combining both networks approaches. The results show that when the CNN uses the RBF to identify and classify sunspots from image processing, its performance is higher than when only CNN is used.
Thiago O. Camargo, Sthefanie Monica Premebida, Denise Pechebovicz, Vinicios R. Soares, Marcella Martins, Virginia Baroncini, Hugo Siqueira, Diego Oliva

Heuristic Method for Optimal Deployment of Electric Vehicle Charge Stations Using Linear Programming

The conventional automobile fleet has significantly increased the emission of toxic gases, thus reducing the quality of air. Therefore, this article proposes a heuristic planning model to promote the massive introduction of plug-in electric vehicles (PEV). Further, this article seeks to deploy electric vehicle charging stations (EVCS), such that the parking time to recharge a PEV are significantly reduced, according to the needs of the user. Besides, the trajectories (driving range) and vehicular flow (traffic) are considered as constraints to the planning problem, which are closely linked to the capacity of the road. On the other hand, clustering techniques are used taking into account real mobility restrictions as a function of minimum distances, and the relationship of the PEV with different charge supply subregions. At last, the model was developed in the Matlab and LpSolve environments. The former will enable the analysis of different trajectories and their relationship with its surroundings. On the other hand, the latter solves the optimization problem using the simplex method.
Miguel Campaña, Esteban Inga

Fuzzy Classification of Industrial Data for Supervision of a Dewatering Machine: Implementation Details and Results

In this document, real data collected in an industrial process are studied and analyzed, with the intention of improving the process supervision seeking for operational efficiency and saving resources, emphasizing in the information cleaning process using basic statistics and data analysis based on non-supervised clustering algorithms: Lamda, GK means and Fuzzy C-means. A general data cleaning procedure for use in industrial environments is suggested. The procedure proposed is followed in a case for a centrifuge machine for mud treatment, three versions of fuzzy classifiers were tested where fuzzy, c-means was finally selected and a result is obtained that permits detecting an inefficient operating state, in some cases the machine was running at a normal current and spending energy and other resources for a long period and the mud was not treated properly, the exit mud was practically the same as the mud at the entrance. The trained classifier has been implemented directly in the PLC used to control the machine, and the results of online classification have been verified showing that states correspond with the process behavior.
Carlos M. Sánchez M, Henry O. Sarmiento M


Weitere Informationen

Premium Partner